[PATCH 0/3 v3] iomap: Use FUA for O_DSYNC DIO writes

[PATCH 0/3 v3] iomap: Use FUA for O_DSYNC DIO writes

[Dave Chinner]

Hi folks,

Version 3 of the FUA for O-DSYNC patchset. This version fixes bugs
found in the previous version. Functionality is otherwise the same
as described in the first version:

https://marc.info/?l=linux-xfs&m=152213446528167&w=2

Version 3:

- fixed O_SYNC behaviour as noticed by Jan Kara
- fixed use after free on IO completion due
  iomap_dio_complete_work() simplification added in version 2. Found
  by KASAN when running xfstests.

Version 2:
- Fixed comment typos in first patch
- simplified iomap_dio_complete_work()
- changed IOMAP_DIO_WRITE_SYNC to IOMAP_DIO_NEED_SYNC
- split blk_queue_fua() into it's own patch
- fixed formatting issue in last patch
- update bio->io_opf directly rather than use bio_set_op_attrs()
- Updated comment to mention we try to use FUA optimistically.

[PATCH 1/3] xfs: move generic_write_sync calls inwards

[Dave Chinner]

From: Dave Chinner <dchinner@redhat.com>

To prepare for iomap iinfrastructure based DSYNC optimisations.

While moving the code araound, move the XFS write bytes metric
update for direct IO into xfs_dio_write_end_io callback so that we
always capture the amount of data written via AIO+DIO. This fixes
the problem where queued AIO+DIO writes are not accounted to this
metric.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
---
 fs/xfs/xfs_file.c | 48 ++++++++++++++++++++++++++++++++---------------
 1 file changed, 33 insertions(+), 15 deletions(-)

diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index bbfbb8f32659..4cbd15d84bb8 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -416,6 +416,12 @@ xfs_dio_write_end_io(
 	if (size <= 0)
 		return size;
 
+	/*
+	 * Capture amount written on completion as we can't reliably account
+	 * for it on submission.
+	 */
+	XFS_STATS_ADD(ip->i_mount, xs_write_bytes, size);
+
 	if (flags & IOMAP_DIO_COW) {
 		error = xfs_reflink_end_cow(ip, offset, size);
 		if (error)
@@ -564,6 +570,11 @@ xfs_file_dio_aio_write(
 	 * complete fully or fail.
 	 */
 	ASSERT(ret < 0 || ret == count);
+
+	if (ret > 0) {
+		/* Handle various SYNC-type writes */
+		ret = generic_write_sync(iocb, ret);
+	}
 	return ret;
 }
 
@@ -601,7 +612,16 @@ xfs_file_dax_write(
 	}
 out:
 	xfs_iunlock(ip, iolock);
-	return error ? error : ret;
+	if (error)
+		return error;
+
+	if (ret > 0) {
+		XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
+
+		/* Handle various SYNC-type writes */
+		ret = generic_write_sync(iocb, ret);
+	}
+	return ret;
 }
 
 STATIC ssize_t
@@ -671,6 +691,12 @@ xfs_file_buffered_aio_write(
 out:
 	if (iolock)
 		xfs_iunlock(ip, iolock);
+
+	if (ret > 0) {
+		XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
+		/* Handle various SYNC-type writes */
+		ret = generic_write_sync(iocb, ret);
+	}
 	return ret;
 }
 
@@ -695,8 +721,9 @@ xfs_file_write_iter(
 		return -EIO;
 
 	if (IS_DAX(inode))
-		ret = xfs_file_dax_write(iocb, from);
-	else if (iocb->ki_flags & IOCB_DIRECT) {
+		return xfs_file_dax_write(iocb, from);
+
+	if (iocb->ki_flags & IOCB_DIRECT) {
 		/*
 		 * Allow a directio write to fall back to a buffered
 		 * write *only* in the case that we're doing a reflink
@@ -704,20 +731,11 @@ xfs_file_write_iter(
 		 * allow an operation to fall back to buffered mode.
 		 */
 		ret = xfs_file_dio_aio_write(iocb, from);
-		if (ret == -EREMCHG)
-			goto buffered;
-	} else {
-buffered:
-		ret = xfs_file_buffered_aio_write(iocb, from);
+		if (ret != -EREMCHG)
+			return ret;
 	}
 
-	if (ret > 0) {
-		XFS_STATS_ADD(ip->i_mount, xs_write_bytes, ret);
-
-		/* Handle various SYNC-type writes */
-		ret = generic_write_sync(iocb, ret);
-	}
-	return ret;
+	return xfs_file_buffered_aio_write(iocb, from);
 }
 
 #define	XFS_FALLOC_FL_SUPPORTED						\
-- 
2.17.0

[PATCH 2/3] iomap: iomap_dio_rw() handles all sync writes

[Dave Chinner]

From: Dave Chinner <dchinner@redhat.com>

Currently iomap_dio_rw() only handles (data)sync write completions
for AIO. This means we can't optimised non-AIO IO to minimise device
flushes as we can't tell the caller whether a flush is required or
not.

To solve this problem and enable further optimisations, make
iomap_dio_rw responsible for data sync behaviour for all IO, not
just AIO.

In doing so, the sync operation is now accounted as part of the DIO
IO by inode_dio_end(), hence post-IO data stability updates will no
long race against operations that serialise via inode_dio_wait()
such as truncate or hole punch.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
---
 fs/iomap.c        | 21 +++++++++++++++------
 fs/xfs/xfs_file.c |  5 -----
 2 files changed, 15 insertions(+), 11 deletions(-)

diff --git a/fs/iomap.c b/fs/iomap.c
index afd163586aa0..b044d8ee2efd 100644
--- a/fs/iomap.c
+++ b/fs/iomap.c
@@ -685,6 +685,7 @@ EXPORT_SYMBOL_GPL(iomap_seek_data);
  * Private flags for iomap_dio, must not overlap with the public ones in
  * iomap.h:
  */
+#define IOMAP_DIO_NEED_SYNC	(1 << 29)
 #define IOMAP_DIO_WRITE		(1 << 30)
 #define IOMAP_DIO_DIRTY		(1 << 31)
 
@@ -759,6 +760,13 @@ static ssize_t iomap_dio_complete(struct iomap_dio *dio)
 			dio_warn_stale_pagecache(iocb->ki_filp);
 	}
 
+	/*
+	 * If this is a DSYNC write, make sure we push it to stable storage now
+	 * that we've written data.
+	 */
+	if (ret > 0 && (dio->flags & IOMAP_DIO_NEED_SYNC))
+		ret = generic_write_sync(iocb, ret);
+
 	inode_dio_end(file_inode(iocb->ki_filp));
 	kfree(dio);
 
@@ -769,13 +777,8 @@ static void iomap_dio_complete_work(struct work_struct *work)
 {
 	struct iomap_dio *dio = container_of(work, struct iomap_dio, aio.work);
 	struct kiocb *iocb = dio->iocb;
-	bool is_write = (dio->flags & IOMAP_DIO_WRITE);
-	ssize_t ret;
 
-	ret = iomap_dio_complete(dio);
-	if (is_write && ret > 0)
-		ret = generic_write_sync(iocb, ret);
-	iocb->ki_complete(iocb, ret, 0);
+	iocb->ki_complete(iocb, iomap_dio_complete(dio), 0);
 }
 
 /*
@@ -961,6 +964,10 @@ iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
 	return copied;
 }
 
+/*
+ * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
+ * is being issued as AIO or not.
+ */
 ssize_t
 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 		const struct iomap_ops *ops, iomap_dio_end_io_t end_io)
@@ -1006,6 +1013,8 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 			dio->flags |= IOMAP_DIO_DIRTY;
 	} else {
 		dio->flags |= IOMAP_DIO_WRITE;
+		if (iocb->ki_flags & IOCB_DSYNC)
+			dio->flags |= IOMAP_DIO_NEED_SYNC;
 		flags |= IOMAP_WRITE;
 	}
 
diff --git a/fs/xfs/xfs_file.c b/fs/xfs/xfs_file.c
index 4cbd15d84bb8..64112a57b004 100644
--- a/fs/xfs/xfs_file.c
+++ b/fs/xfs/xfs_file.c
@@ -570,11 +570,6 @@ xfs_file_dio_aio_write(
 	 * complete fully or fail.
 	 */
 	ASSERT(ret < 0 || ret == count);
-
-	if (ret > 0) {
-		/* Handle various SYNC-type writes */
-		ret = generic_write_sync(iocb, ret);
-	}
 	return ret;
 }
 
-- 
2.17.0

[PATCH 3/3] iomap: Use FUA for pure data O_DSYNC DIO writes

[Dave Chinner]

From: Dave Chinner <dchinner@redhat.com>

If we are doing direct IO writes with datasync semantics, we often
have to flush metadata changes along with the data write. However,
if we are overwriting existing data, there are no metadata changes
that we need to flush. In this case, optimising the IO by using
FUA write makes sense.

We know from the IOMAP_F_DIRTY flag as to whether a specific inode
requires a metadata flush - this is currently used by DAX to ensure
extent modification as stable in page fault operations. For direct
IO writes, we can use it to determine if we need to flush metadata
or not once the data is on disk.

Hence if we have been returned a mapped extent that is not new and
the IO mapping is not dirty, then we can use a FUA write to provide
datasync semantics. This allows us to short-cut the
generic_write_sync() call in IO completion and hence avoid
unnecessary operations. This makes pure direct IO data write
behaviour identical to the way block devices use REQ_FUA to provide
datasync semantics.

On a FUA enabled device, a synchronous direct IO write workload
(sequential 4k overwrites in 32MB file) had the following results:

# xfs_io -fd -c "pwrite -V 1 -D 0 32m" /mnt/scratch/boo

kernel		time	write()s	write iops	Write b/w
------		----	--------	----------	---------
(no dsync)	 4s	2173/s		2173		8.5MB/s
vanilla		22s	 370/s		 750		1.4MB/s
patched		19s	 420/s		 420		1.6MB/s

The patched code clearly doesn't send cache flushes anymore, but
instead uses FUA (confirmed via blktrace), and performance improves
a bit as a result. However, the benefits will be higher on workloads
that mix O_DSYNC overwrites with other write IO as we won't be
flushing the entire device cache on every DSYNC overwrite IO
anymore.

Signed-Off-By: Dave Chinner <dchinner@redhat.com>
Reviewed-by: Christoph Hellwig <hch@lst.de>
---
 fs/iomap.c | 51 ++++++++++++++++++++++++++++++++++++++++++++++-----
 1 file changed, 46 insertions(+), 5 deletions(-)

diff --git a/fs/iomap.c b/fs/iomap.c
index b044d8ee2efd..bcfd7f3654d4 100644
--- a/fs/iomap.c
+++ b/fs/iomap.c
@@ -685,6 +685,7 @@ EXPORT_SYMBOL_GPL(iomap_seek_data);
  * Private flags for iomap_dio, must not overlap with the public ones in
  * iomap.h:
  */
+#define IOMAP_DIO_WRITE_FUA	(1 << 28)
 #define IOMAP_DIO_NEED_SYNC	(1 << 29)
 #define IOMAP_DIO_WRITE		(1 << 30)
 #define IOMAP_DIO_DIRTY		(1 << 31)
@@ -861,6 +862,7 @@ iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
 	struct iov_iter iter;
 	struct bio *bio;
 	bool need_zeroout = false;
+	bool use_fua = false;
 	int nr_pages, ret;
 	size_t copied = 0;
 
@@ -884,8 +886,20 @@ iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
 	case IOMAP_MAPPED:
 		if (iomap->flags & IOMAP_F_SHARED)
 			dio->flags |= IOMAP_DIO_COW;
-		if (iomap->flags & IOMAP_F_NEW)
+		if (iomap->flags & IOMAP_F_NEW) {
 			need_zeroout = true;
+		} else {
+			/*
+			 * Use a FUA write if we need datasync semantics, this
+			 * is a pure data IO that doesn't require any metadata
+			 * updates and the underlying device supports FUA. This
+			 * allows us to avoid cache flushes on IO completion.
+			 */
+			if (!(iomap->flags & (IOMAP_F_SHARED|IOMAP_F_DIRTY)) &&
+			    (dio->flags & IOMAP_DIO_WRITE_FUA) &&
+			    blk_queue_fua(bdev_get_queue(iomap->bdev)))
+				use_fua = true;
+		}
 		break;
 	default:
 		WARN_ON_ONCE(1);
@@ -933,10 +947,14 @@ iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
 
 		n = bio->bi_iter.bi_size;
 		if (dio->flags & IOMAP_DIO_WRITE) {
-			bio_set_op_attrs(bio, REQ_OP_WRITE, REQ_SYNC | REQ_IDLE);
+			bio->bi_opf = REQ_OP_WRITE | REQ_SYNC | REQ_IDLE;
+			if (use_fua)
+				bio->bi_opf |= REQ_FUA;
+			else
+				dio->flags &= ~IOMAP_DIO_WRITE_FUA;
 			task_io_account_write(n);
 		} else {
-			bio_set_op_attrs(bio, REQ_OP_READ, 0);
+			bio->bi_opf = REQ_OP_READ;
 			if (dio->flags & IOMAP_DIO_DIRTY)
 				bio_set_pages_dirty(bio);
 		}
@@ -966,7 +984,12 @@ iomap_dio_actor(struct inode *inode, loff_t pos, loff_t length,
 
 /*
  * iomap_dio_rw() always completes O_[D]SYNC writes regardless of whether the IO
- * is being issued as AIO or not.
+ * is being issued as AIO or not.  This allows us to optimise pure data writes
+ * to use REQ_FUA rather than requiring generic_write_sync() to issue a
+ * REQ_FLUSH post write. This is slightly tricky because a single request here
+ * can be mapped into multiple disjoint IOs and only a subset of the IOs issued
+ * may be pure data writes. In that case, we still need to do a full data sync
+ * completion.
  */
 ssize_t
 iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
@@ -1012,10 +1035,21 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 		if (iter->type == ITER_IOVEC)
 			dio->flags |= IOMAP_DIO_DIRTY;
 	} else {
+		flags |= IOMAP_WRITE;
 		dio->flags |= IOMAP_DIO_WRITE;
+
+		/* for data sync or sync, we need sync completion processing */
 		if (iocb->ki_flags & IOCB_DSYNC)
 			dio->flags |= IOMAP_DIO_NEED_SYNC;
-		flags |= IOMAP_WRITE;
+
+		/*
+		 * For datasync only writes, we optimistically try using FUA for
+		 * this IO.  Any non-FUA write that occurs will clear this flag,
+		 * hence we know before completion whether a cache flush is
+		 * necessary.
+		 */
+		if ((iocb->ki_flags & (IOCB_DSYNC | IOCB_SYNC)) == IOCB_DSYNC)
+			dio->flags |= IOMAP_DIO_WRITE_FUA;
 	}
 
 	if (iocb->ki_flags & IOCB_NOWAIT) {
@@ -1071,6 +1105,13 @@ iomap_dio_rw(struct kiocb *iocb, struct iov_iter *iter,
 	if (ret < 0)
 		iomap_dio_set_error(dio, ret);
 
+	/*
+	 * If all the writes we issued were FUA, we don't need to flush the
+	 * cache on IO completion. Clear the sync flag for this case.
+	 */
+	if (dio->flags & IOMAP_DIO_WRITE_FUA)
+		dio->flags &= ~IOMAP_DIO_NEED_SYNC;
+
 	if (!atomic_dec_and_test(&dio->ref)) {
 		if (!is_sync_kiocb(iocb))
 			return -EIOCBQUEUED;
-- 
2.17.0

Re: [PATCH 0/3 v3] iomap: Use FUA for O_DSYNC DIO writes

[Darrick J. Wong]

On Wed, May 02, 2018 at 03:38:04PM +1000, Dave Chinner wrote:
> Hi folks,
> 
> Version 3 of the FUA for O-DSYNC patchset. This version fixes bugs
> found in the previous version. Functionality is otherwise the same
> as described in the first version:
> 
> https://marc.info/?l=linux-xfs&m=152213446528167&w=2

Will test the whole series, consider this a tentative:
Reviewed-by: Darrick J. Wong <darrick.wong@oracle.com>

--D

> 
> Version 3:
> 
> - fixed O_SYNC behaviour as noticed by Jan Kara
> - fixed use after free on IO completion due
>   iomap_dio_complete_work() simplification added in version 2. Found
>   by KASAN when running xfstests.
> 
> Version 2:
> - Fixed comment typos in first patch
> - simplified iomap_dio_complete_work()
> - changed IOMAP_DIO_WRITE_SYNC to IOMAP_DIO_NEED_SYNC
> - split blk_queue_fua() into it's own patch
> - fixed formatting issue in last patch
> - update bio->io_opf directly rather than use bio_set_op_attrs()
> - Updated comment to mention we try to use FUA optimistically.
> 
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-xfs" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html